Related Experiment Video
Updated: May 5, 2026
08:20
MRI-guided dmPFC-rTMS as a Treatment for Treatment-resistant Major Depressive Disorder
Published on: August 11, 2015
13.0K
3DViT-GAT: a unified atlas-based 3D vision transformer and graph learning framework for major depressive disorder
Nojod M Alotaibi1, Areej M Alhothali2, Manar S Ali2
1Department of Computer Science, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah, Saudi Arabia. nalotaibi0351@stu.kau.edu.sa.
Scientific Reports
|March 2, 2026
Summary
This study introduces a novel deep learning approach using Vision Transformers and Graph Neural Networks for detecting major depressive disorder (MDD) from brain scans. Atlas-based methods showed superior performance, highlighting the value of anatomical priors in diagnosing MDD.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Psychiatry
Background:
- Major Depressive Disorder (MDD) is a widespread mental health issue impacting global health.
- Current automated detection methods for MDD using structural magnetic resonance imaging (sMRI) often rely on limited feature extraction techniques.
- There is a need for advanced deep learning (DL) models that can capture complex brain patterns for improved MDD diagnosis.
Purpose of the Study:
- To develop and evaluate a unified DL pipeline for automated MDD detection using sMRI data.
- To compare the efficacy of an atlas-based region definition strategy against a cube-based approach within the pipeline.
- To leverage Vision Transformers (ViTs) for feature extraction and Graph Neural Networks (GNNs) for classification of MDD.
Main Methods:
- A novel pipeline integrating ViTs for 3D region embedding extraction from sMRI and GNNs for classification was developed.
- Two region definition strategies were explored: an atlas-based approach using predefined brain atlases and a cube-based method.
- Cosine similarity graphs were constructed to model inter-regional relationships, guiding the GNN classification.
Main Results:
- The proposed model achieved high performance on the REST-meta-MDD dataset, with the best model yielding 81.51% accuracy via 10-fold cross-validation.
- Specific performance metrics included 85.94% sensitivity, 76.36% specificity, 80.88% precision, and 83.33% F1-score.
- The atlas-based models consistently outperformed the cube-based approach, demonstrating the advantage of incorporating anatomical priors.
Conclusions:
- The developed ViT-GNN pipeline offers a promising approach for automated MDD detection using sMRI.
- Utilizing predefined brain atlases in the region definition strategy is crucial for enhancing diagnostic accuracy in MDD detection.
- This study underscores the importance of integrating domain-specific anatomical knowledge into DL models for psychiatric disorder diagnosis.
